agent coordination - kth
TRANSCRIPT
ID2209 Distributed Artificial Intelligence and Intelligent Agents
Agent Coordination
Mihhail Matskin: www.ict.kth.se/courses/ID2209
Autumn 2015
Agent Coordination • general models
– why? – how? – effective control of distributed search
• common coordination techniques – organizational structures – meta-level information exchange – multi-agent planning – explicit analysis and synchronization – coordination based on human teamwork – commitments and conventions
References - Curriculum
• Wooldridge: ”Introduction to MAS”,
– Chapter 8
Coordination Example
Consider an interaction between two robots, A and B,
operating in a warehouse. The robots have been designed
by different companies, and they are stacking and
unstacking boxes to remove certain goods that have been
stored in the building. They need to coordinate their
actions to share the work load and to avoid knocking
into each other and dropping the boxes.
Another Coordination Problem
• Managing the
interdependencies between the
activities of agents. e.g.
– You and I both want to leave the
room. We independently walk
towards the door, which can only
fit one of us. I graciously permit
you to leave first.
Cartoon taken from Klein, AAMAS2002.
Coordination
”The process by which an agent reasons about its
local actions and the (anticipated) actions of
others to try and ensure that the community acts
in a coherent manner.”
Jennings,1996
Coordination
• Preventing anarchy or chaos • Meeting global constraints • Distributed expertise, resources or information • Dependencies between agents’ actions • Efficiency
Coordination To ensure: • coverage • connectivity • rationality • capability
Activities: • supplying timely information to needy agents • ensuring synchronization • avoiding redundant problem solving
How to reach coherent behavior? • complete knowledge • centralized control • distributed control
Fundamental coordination processes
• mutual adjustment • direct supervision • standardization
Basic steps in coordination problem
• analysis of a perspective for coordination – external observation may be not enough
• coordinated but incoherent • coherent but non-coordinated
– examining an internal goals and motivations is necessary
• applying appropriate coordination model
Control decisions and search View to coordination as a matter of effective control of
distributed search. What is a control? • control decisions are decisions about what actions to take next • control decisions are choices • control knowledge is any knowledge that informs control
decisions • each control choice is the outcome of an overall control regime
which includes knowledge about: – what are the control alternatives? – what are the decision criteria? – what is the decision procedure?
Control in cooperative distributed problem solving
Control choice becomes more complicated to the
degree that control decisions are more
• numerous • asynchronous • decentralized
DAI and distributed search • the space of alternative problem states (goals) can be seen as a
large search space investigated by a number of agents • each agent has to make local control decisions • these local control decisions have impacts on the overall
efforts done by the collection of problem solvers
Search space
1
2 3a 3b
goal
Control uncertainty Control decision uncertainty - ambiguity in the next
action choice (often is characterized as the size of the set of next-states)
A
B C
Two kinds of control in DAI
• network control or cooperative control - comprises decision procedures that lead to good overall performance
• local control - refers to decision procedures that lead to good local decisions, and that are based on local information only
Network control Sets contexts for agent's individual control decisions based on
network level information
Network-level information is: • aggregated from more than one agent or abstracted from data
from more than one agent • information that concerns the relationships among a collection
of agents
Network-level information can be utilized to influence: • set of action alternatives to consider a control decisions • the decision criteria applied to choose one of them • the control decision procedure
Network control One type of network control (or coordination) is
allocation of search-space regions to agents • search space - alternative problem states (goals) • search process explores that space by generating a tree of
possibilities • since a search process explores a tree of possibilities (and any
tree is recursively composed of subtrees) it follows that any region of the search space can be characterized by a set of subtree roots
• allocation of search-space regions to agents takes place by allocating collections of search-subtree roots to agents
• dynamic nature of region allocation
Local control
Concerns the status and process of a single node in its own local environment and its own local search-space region
Interaction with network control: • network may increase or decrease local control
uncertainty • better local control may more efficiently
uncover information that can focus network control decision.
Reducing uncertainty Reducing degree and/or reducing impact of uncertainty
– Impact - arbitrariness of control decisions – Impact of control uncertainty can be reduced by reducing
common dependencies that agents share
Two kinds of dependencies: – logical dependencies – resource dependencies
A very important part of coordination is analysis of interdependencies
Search tree (example) G0
G1 G2
G3 G4 G5 G7
G8 G9 G10 G11
agent1 agent2
G9 ∨ G10
G1 & G2
and
or
resources r1 r2
Joint goals • team members are mutually responsible to one another • the team members have a joint commitment to the joint
activity • the team members are committed to be mutually supportive of
one another
G1 G2
G3 G4 G5
G6 G7
agent1 agent2
Activities in coordination
• defining the goal graph • assigning regions of search space • controlling decisions about which areas of the graph
to explore • traversing the goal structure satisfying dependencies • ensuring report of the successful traversal
Determining the approach for each of the phases is a matter of system design and it depends upon:
• the nature of the domain • the type of agents included into community • the desired solution characteristics
Coordination Techniques
• Organisational Structures
• Meta-level Information Exchange
– e.g. Partial Global Planning (PGP), (Durfee)
• Multi-agent Planning
• Explicit Analysis and Syncronization
• Norms and social laws
• Coordination Models based on human teamwork:
– Joint commitments (Jennings)
– Mutual Modelling
Comparing Common Coordination Techniques
Organisation structures
Meta-level Information exchange
Multi-agent Planning
low low less
high high more
P r e d i c t a b i l i t y
R e a c t i v t y
I n f o E x c h a n g e
Explicit analysis and synchronization
Organizational structures • the simplest coordination approach • implicit coordination • provides framework for defining
– roles – communication paths – authority relationships
• pre-defined, long-term relations • specifies the distribution of specializations among
agents • a precise way of dividing the problem space without
specifying particular problems • agents are associated with problem types and problem
instances circulate to the agents which are responsible for instances of that type
Organizational Structures
• Organisational structures may be:
– Functional
– Spatial
– Product-Oriented
Organizational Structure Models
• A pattern for decision-making and communication among
a set of agents who perform tasks in order to achieve goals.
e.g.
– Automobile industry
• Has a set of goals: To produce different lines of cars
• Has a set of agents to perform the tasks: designers, engineers,
salesmen • Costs considered
– production costs – coordination costs – vulnerability costs
Reference: Malone 1987
Alternative Coordination Structures 1 Product Hierarchy
Designer
Product Manager I
Salesman Engineer Designer
Product Manager 2
Salesman Engineer Designer
Product Manager 3
Salesman Engineer
Product hierarchy
• several divisions for different product lines • each division has a product manager • each division has its own separate departments (agents) for
different functions • product manager assigns tasks to agents in its department • communication links in product structure • fails inside department do not affect other products • one message to assign task and one message to notify result • also model for set of separate companies
Product Manager (several products)
Alternative Coordination Structures 2 Functional Hierarchy
Designers
Design Manager
Salesmen
Sales Manager
Engineers
Engineering Manager
Functional hierarchy
• a number of agents of similar types are pooled into functional departments
• each department has a functional manager • reduce duplication of efforts • executive office - production manager for all products
hierarchical task allocation • 2 messages to assign task and 2 messages to notify about
results • failure of task agent - delay and task reallocation • managers are critical
Markets
• not necessary to make all components by itself
• subcontracting
Alternative Coordination Structures 4 Decentralised Market
Product Manager 2
Designers Salesmen Engineers
Product Manager 1 Product Manager 3
Decentralized market
• all buyers are in contact with all possible suppliers • buyers play a role of product managers • each has a communication link with each supplier • buyers can choose the best supplier • m suppliers - 2m+2 messages • processor fails the task is reassigned • if manager fails – product is failed
Alternative Coordination Structures 3 Centralised Market
Product Manager 2
Designers
Design Manager
Salesmen
Sales Manager
Engineers
Engineering Manager
Product Manager 1 Product Manager 3
Functional Managers
Centralized market
• brokers are in contact with possible sellers • fewer connections and communications are required • brokers play a role of functional managers • broker chooses the best supplier • 4 messages • similar to functional model • failure of one product manager does not affect others
Summary of organization structures Processors shared among products
Messages required to assign task
Result of task processor failure
Result of functional manager failure
Result of product manager failure
Product hierarchy Decentralised market Centralised market Funtional hierarchy
no 2 1 product disrupted - 1 prod.
disrupted
yes 2m+2 Task reassigned - 1 prod.
disrupted
yes 4 Task reassigned
All prod. disrupted
1 prod. disrupted
yes 4 Task reassigned
All prod. disrupted
All prod. disrupted
Comparison of Organization Structures
Production cost
Coordination cost
Vulnerability cost
Product hierarchy
H L H
Functional hierarchy
L M- H+
Centralised market
L M+ H-
Decentralised market
L H L
Changes in costs as size of structure increases
Production cost
Coordination cost
Vulnerability cost
Product hierarchy
+ + +++
Funtional hierarchy
+ ++ +
Centralised market
+ +++ ++
Decentralised market
+ ++++ ++
Organizational Structures - Critique
• Useful when there are master/slave relationships in the
MAS.
• Control over the slaves actions – mitigates against benefits
of DAI such as reliability, concurrency.
• Presumes that atleast one agent has global overview – an
unrealistic assumption in MAS.
Comparing Common Coordination Techniques
Organisation structures
Meta-level Information exchange
Multi-agent Planning
low low less
high high more
P r e d i c t a b i l i t y
R e a c t i v t y
I n f o E x c h a n g e
Explicit analysis and synchronization
Meta-level Information Exchange
• Exchange control level information about current priorities and focus.
• Control level information – May change
– Influence the decisions of agents
• Does not specify which goals an agent will or will not consider.
• Imprecise
• Medium term – can only commit to goals for a limited amount of time.
Partial Global Planning (PGP) Durfee
• The basic condition - requirement that several distributed agents work on solution of the same problem
• An agent can observe actions and relations between groups of other agents
A B
A1 A2 A3
A B
B1 B2 A2
Agent 1 Agent 2
communication
Partial Global Planning (PGP)
• PGP involves 3 iterated stages:
1. Each agent decides what its own goals are and
generates short-term plans in order to achieve them.
2. Agents exchange information to determine where
plans and goals interact.
3. Agents alter local plans in order to better coordinate
their own activities.
Partial Global Planning (PGP) • Task decomposition • Local plan formulation • Local plan abstraction • Communication of Local plan abstraction • Partial global goal identification • PGP construction and modification • Communication planning • Acting on PGP • Ongoing modifications • Task reallocation
Partial Global Planning (PGP) 1
• A DAI testbed – Distributed Vehicle Monitoring Testbed
(DVMT) – to successfully track a number of vehicles that
pass within the range of a set of distributed sensors (agents).
• Geographically distributed agents • Each agent monitors sensing a portion
of the overal area
• Discrete time-interval
• By analyses of sound – type of vehicle and movement map
• Provides information to another agent
• There should be overlapping areas Overlapping
area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Partial Global Planning (PGP)
• Main principle: cooperating agents exchange information
in order to reach common conclusions about the problem
solving process.
• Why is planning partial?
– A agent does not generate a plan for the entire problem.
• Why is planning global?
– Agents form non-local plans by exchanging local plans and
cooperating to achieve a non-local view of problem solving.
Task decomposition (PGP) • Starts with the premise that tasks
are inherently decomposed.
• Assumes that an agent with a task to plan for might be unaware of what tasks other agents might be planning for and how those tasks are related to its own.
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
• No individual agent might be aware of the global tasks or states.
• Purpose of coordination is to develop sufficient awareness.
Local plan formulation (PGP) • before agent communicates with
other it must first develop understanding of what goals it is trying to achieve and which actions to perform
• local plans will most often be uncertain, involving branches of alternative actions depending on results of previous actions and changes in environments
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Local plan abstraction (PGP) • alternative courses of action for
achieving the same goal are important for an agent
• however, the detail of alternatives might be unnecessary considering the agent's ability to coordinate with others
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
• an agent might have to commit to activities at one level of details without committing to activities at more detailed levels
• agents are designed to identify their major plan steps that could be of interest to other agents
Communication (PGP)
• agents must communicate about abstract local plans in order to build models of joint activity
• in PGP, the knowledge to guide this communication is contained in the meta-level-organization
Partial global goal identification (PGP)
• the exchange of local plans gives an opportunity to identify when the goal of one or more agents could be subgoals of a single global goal
• only portions of the global goal might be known to agents
• partial global goal
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
PGP construction and modification
• local plans can be integrated into PGP
• PGP can identify opportunities for improved coordination (performing related tasks earlier, avoiding redundant task achievement)
• reordering of actions (usage of action rates)
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
PGP construction and modification Rates: • whether the task is unlikely to
have been accomplished already by another agent
• how long it is expected to take • how useful its results will be to
others in performing their tasks
Algorithm 1. For the current ordering, rate the individual actions and sum the ratings 2. For each action, examine the later actions for the same agent and find
the most highly-rated one. If it is higher rated, then swap the actions 3. If the new ordering is more highly rated than the current one, then
replace the current ordering with the new one and go to step 2. 4. Return the current ordering
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Communication planning (PGP) • agent must next consider what
interactions should take place between agents
• interactions in form of communicating the results of tasks are planned
• by examining PGP an agent can determine when a task will be completed by one agent that could be of interest to another agent and can explicitly plan the communication action to transmit the result
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Communication planning (PGP) 1. Initialize the set of partial task
results to integrate 2. While the set contains more than
one element: (a) For each pair of elements: find the
earliest time and agent at which they can be combined
(b) For the pair that can be combined earliest: add a new element to the set of partial results for the combination and remove the two elements that were combined
3. Return the single element in the set
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Acting on PGP • once a PGP has been constructed
and the concurrent local and communication actions have been ordered, the collective activities of the agents have been planned
• these activities must be translated back to the local level
• an agent responds to a change in its PGP by modifying the abstract representation of its local plans
• modified representation is used by agent when choosing its next local action
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Ongoing modifications (PGP)
• as agents pursue their plans, their actions or events in environment might lead to changes in tasks
• a change in coordination is deciding when the changes in local plans are significant enough
• defining a threshold (significant temporal deviations)
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Task reallocation (PGP)
• disproportional task load
• through PGP exchange of abstract models of agent's activities and detection whether they are overburdened
• possible task reallocation
Overlapping area
Agenti
Agentj
Vehicle track
1
2
3
4
5
6
7
8
Partial Global Planning (PGP) • Partial Global Plan: a cooperatively generated
datastructure containing the actions and interactions of a
group of agents.
• Contains:
– Objective – the larger goal of the system.
– Activity map – what agents are actually doing and the results
generated by the activities.
– Solution construction graph – a representation of how the agents
ought to interact in order to successfully generate a solution.
Summary (PGP) • PGP particularly suited to applications where
some uncoordinated activity can be tolerated and overcome
• PGP focuses on dynamically revising plans in cost-effective ways given an uncertain world
• PGP works well for many tasks, but could be inappropriate for domains such as air-traffic control where guarantees about coordination must be made prior to any execution
Coordination Techniques
• Organisational Structures
• Meta-level Information Exchange
– e.g. Partial Global Planning (PGP), (Durfee)
Ø Multi-agent Planning
• Norms and social laws
• Coordination Models based on human teamwork:
– Joint commitments (Jennings)
– Mutual Modelling
Comparing Common Coordination Techniques
Organisation structures
Meta-level Information exchange
Multi-agent Planning
low low less
high high more
P r e d i c t a b i l i t y
R e a c t i v t y
I n f o E x c h a n g e
Explicit analysis and synchronization
Multi-agent Planning • Agents generate, exchange and synchronize explicit plans
of actions to coordinate their joint activity.
• They arrange apriori precisely which tasks each agent will take on.
• Plans specify a sequence of actions for each agent.
• Consideration of uncertainty is moved to the planning activity
• More specific than organizational structures or PGP and shorter time-horizon
Multi-agent Planning
• Two basic approaches:
1. Centralised – a central coordinator develops,
decomposes and allocates plans to individual agents.
2. Distributed – a group of agents cooperate to form:
1. Centralised plan
2. Distributed plan
Multi-agent Planning Centralized planning for distributed plans
• Given a goal description, a set of operators and an initial state description, generate a partial order plan
• Decompose the plan into subproblems such that ordering relationships between steps tend to be minimized across subproblems
• Insert synchronization actions into subplans • Allocate subplans to agents using task-passing
mechanisms • Initiate plan execution and optionally monitor
progress
Multi-agent Planning • Distributed Planning for centralised plans:
– e.g. Air traffic control domain (Cammarata) • Aim: Enable each aircraft to maintain a flight plan that will
maintain a safe distance with all aircrafts in its vicinity. • Each aircraft sends a central coordinator information about its
intended actions. The coordinator builds a plan which specifies all of the agents’ actions including the ones that they should take to avoid collision.
– e.g. A general technological process (manufacturing) • Centralized controll and plan is centralized but each step can
be planned in parallel
• Disitrbuted expertize but control is centralized
Multi-agent Planning
• Distributed Planning for distributed plans: – Individual plans of agents, coordinated dynamically.
– No individual with a complete view of all the agents’
actions.
– More difficult to detect and resolve undesirable
interactions.
Centralized multi-agent plans (plan coordination) - an example (Georgeff)
• robots world • actions - sequences of states • each action has pre- and post-conditions • plans are generated separately
This approach has several steps • interaction analysis • safety analysis • interaction resolution
Centralized multi-agent plans An action a is a sequence of states p : a = {p1, ... , pi, ... , pn}, where p1 is action's initial state, pn is
action’s final state and p2,…, pn-1 are intermediate states
A planning problem consists of • set of states S • a designated set of initial states I • a set of actions A • a set of goal states G
A single-agent plan P is a description of sequence of actions a1, a2, ... , an from A such that:
• a1 is applicable to all initial states (a1 initial state is in I) • for all i, 1 < i ≤ n, the action ai is applicable to all states in the
range of ai-1 • an achieves the goal G (final state of an is in G)
Interaction analysis • establishing incompatible situations • <p, q> - the situation where both p and q hold (it can
be unsatisfiable if contradictory) • possible alternatives for safe action execution
– actions can be executed in parallel (actions said to be commuted)
– suspending one of the action • precedence may exist of one over another (post-
condition of one is a pre-condition for another) – both actions may have precedence over each other – neither actions may have precedence
• plans are generated separately
Safety analysis
Determining unsafe situations for each pair of actions in two plans
• if ai and bj do not commute, then <begin(ai),begin(bj)> is unsafe
• if ai doesn't have precedence over bj, then <begin(ai),end(bj-1)> is unsafe
The set of all such unsafe situations is called interaction set
Rules for safety governing
• if s=<begin(ai),begin(bj)>, then s is unsafe if either of its successor situations are unsafe
• if s=<begin(ai),end(bj)>, then s is unsafe if <end(ai),end(bj)> is unsafe
• if s=<end(ai),end(bj)>, then s is unsafe if both successor situations are unsafe
• together with those situations occurring in the interaction set, these are all the unsafe situations
Interaction resolution
The set of unsafe situations is next analyzed to identify contiguous sequences of unsafe situations
Handling synchronization
• critical regions • communication commands
– P!s - send s to process P – P?s - receive s from P
Centralized multi-agent plans (plan coordination)
• drop commuting actions • exponential search of interleavings • introducing synchronization
Example - 2 robots The actions are informally as follows:
• a1m: agent 1 moves to the lathe • a2m: agent 2 moves to the lathe • a1p: agent 1 places metal stock in lathe • a2p: agent 2 places metal stock in lathe • a1b: agent 1 makes a bolt • a2n: agent 2 makes a nut • a1f: agent 1 moves to end • a2f: agent 2 moves to end
Preconditions for a1b and a2n - the lathe must be in possession of the appropriate agent, as well as postconditions for a1p and a2p
Example (cont.) Assume the simple planner produces the following
single-agent plans:
a1m -> a1p -> a1b -> a1f a2m -> a2p -> a2n -> a2f
The following properties can be established • actions a1b and a2n conflict with one another • actions a1p and a2p each have precedence over each
other but do not commute • action a1b has precedence over a2p, but not vice
versa • action a2n has precedence over a1p, but not vice
versa
Example (cont.) Interaction set is determined to be:
<begin(a1b),begin(a2n)>, <begin(a1b),end(a2p)>,
<end(a1p),begin(a2n)>, <begin(a1p),begin(a2p)>,
<begin(a1b),begin(a2p)>, <end(a1p),begin(a2p)>,
<begin(a1p),begin(a2n)>, <begin(a1p),end(a2p)>
We next form the simplified solutions: begin(a1p) -> end(a1p) -> begin(a1b) -> end(a1b)
begin(a2p) -> end(a2p) -> begin(a2n) -> end(a2n)
Example (cont.) Safety analysis - we get 2 critical regions
begin(a1p) -> end(a1b) begin(a2p) -> end(a2n)
Next we insert synchronization commands into initial plans:
a1m -> S!begin(a1p) -> a1p -> a1b -> S!end(a1b) -> a1f a2m -> S!begin(a2p) -> a2p -> a2n -> S!end(a2n) -> a2f
Synchronizer (very generally):
N, M := false not N; P?begin(a1p) -> M := true not M; Q?begin(a2p) -> N := true true; P?end(a1b) -> M := false true; Q?end(a2n) -> N := false
Multi-agent Planning
• Critique:
– Agents share and process a huge amount of information.
– Requires more computing and communication resources.
• Difference between multi-agent planning and PGP:
– PGP does not require agents to reach mutual agreements
before they start acting.
Comparing Common Coordination Techniques
Organisation structures
Meta-level Information exchange
Multi-agent Planning
low low less
high high more
P r e d i c t a b i l i t y
R e a c t i v t y
I n f o E x c h a n g e
Explicit analysis and synchronization
Explicit analysis and synchronization • analysis of a situation in each decision-making step
(possible-next-action-set) • interacting with other agents
– exchange among all interdependent agents – locks action – uses reply information to prune its next-action set – send synchronization unlocking messages
• a lot of time for communication during planning • if the level of dependency is low and the granularity
of actions is high - can provide useful coordination • short time-horizon
Coordination Techniques
• Organisational Structures
• Meta-level Information Exchange
– e.g. Partial Global Planning (PGP), (Durfee)
• Multi-agent Planning
Ø Norms and social laws
• Coordination Models based on human teamwork:
– Mutual Modelling
– Joint commitments
Social Norms and Laws • Norm: an established, expected pattern of behaviour.
– e.g. To queue when waiting for the bus
• Social laws: similar to Norms, but carry some authority.
– e.g. Traffic rules.
• Social laws in an agent system can be defined as a set of constraints:
– Constraint => 〈E’,α 〉,
• E’ ⊆ E is a set of environment states
• α ∈ Ac is an action, (Ac is the finite set of actions possible for an agent)
Ø if the environment is in some state e ∈ E’, then the action α is forbidden.
An agent (or plan) is said to be legal wrt a social law if it never attempts to
perform an action that is forbidden by some constraint in the social law
Coordination Techniques
• Organisational Structures
• Meta-level Information Exchange
– e.g. Partial Global Planning (PGP), (Durfee)
• Multi-agent Planning
• Norms and social laws
Ø Coordination Models based on human teamwork:
– Mutual Modelling
– Joint commitments
Mutual Modelling
• Build a model of the other agents – their
beliefs and intentions.
Ø Put ourselves in the place of the other
• Coordinate own activities based on this
model
• Cooperation without communication
Coordination, Cooperation & Teamwork
• Can we have coherent behaviour without
cooperation?
”A group of people are sitting in a park. As a
result of a sudden downpour, all of them run to
a tree in the middle of the park because it is the
only source of shelter.”
Teamwork Example
Two vehicles travelling in a convoy:
Consider two agents Bob and Alice. Bobs wants to drive
home, but does not know his way. He knows that Alice is
going near there and that she does know the way. Bob
talks to Alice and they both agree that he follows her
through traffic and that they drive together.
Ref: Cohen & Levesque, 1991
Teamwork
• Important distinction:
– Coherent behaviour that is not cooperative, e.g
• Individual drivers in traffic following traffic rules
– Coordinated cooperative action, e.g
• A convoy of drivers
Teamwork Definition
• American Heritage Dictionary
– Cooperative effort by the members of a
team to achieve a common goal.
Teamwork
• How does an individual intention towards a particular
goal differ from being a part of a team with a
collective intention towards a goal?
– Responsibility towards the other members of the team.
• e.g. You and I are lifting a heavy object
– Agents i, j and k are a team and have a common goal G. G
g2 g3 g1
i j k
Joint Intentions model
• Based on human teamwork models
”When a group of agents are engaged in a
cooperative activity, they must have a joint
commitment to the overall aim as well as their
individual commitments.”
Commitments
• Commitment – a pledge or promise (e.g. to lift the heavy
object).
– Commitment persists – if an agent adopts a commitment, it is not
dropped until for some reason it becomes redundant.
– Commitments may change over time, e.g. due to a change in the
environment
– Main problem with joint commitment:
• Hard to be aware of each others states at all times
Conventions
• Convention – means of monitoring a commitment
– e.g. specifies under what circumstances a commitment can be
abandoned.
• Need conventions to describe when to change a
commitment:
1. When to keep a commitment (retain)
2. When to revise a commitment (rectify)
3. When to remove a commitment (abandon)
Convention - Example
• Reasons for terminating a Commitment: – Commitment Satisfied
– Commitment Unattainable
– Motivation for commitment no longer present
• Rule R1: If Commitment Satisfied OR
Commitment Unattainable OR
Motivation for Commitment no longer present
then
terminate Commitment.
Joint Commitments
• Joint action by a team involves more than just the union of simultaneous individual actions.
• When a group of agents are engaged in a cooperative activity, they must have: • Joint commitment to the overall activity
• Individual commitment to the specific task that they have been assigned to
G
g2 g3 g1
i j k
Social Conventions
• Individual Conventions describe how an agent should monitor its
commitments, but not how it should behave towards other
agents.
– Asocial
– Sufficient for goals that are independent.
• For inter-dependent goals:
– Need social conventions
• Specify how to behave with respect to the other members of the team.
Social Conventions
• Team members must be aware of the convention that govern their interactions. e.g.
G
g1 g2 AND
Ai Aj
G
g1 g2 OR
Ai Aj
• Both Ai and Aj must fulfill their commitments
to achieve G.
• Either Ai or Aj must fulfill their commitment.
Ø There is a need for all agents in a team to
inform other members of the status of their
commitments!
Coordination Summary • Coordination – ensuring coherent behaviour.
• Can be considered as a control over distributed search
• Coordination Techniques: – Organisational structures
– Meta-level information exchange
– Multi-agent Planning
– Social norms and laws
– Mutual Modelling
– Joint Intentions
References – Recommended Reading for Teamwork
• Not curriculum:
– Cohen, P. R. and Levesque, H. J., ”Teamwork”, Nous, 25, 1991.
– Tambe, M., ”Towards Flexible Teamwork”, Journal of Artificial
Intelligence Research, Volume 7, 1997, pp. 83-124.
Next Topic: MAS Architectures Wooldridge: ”Introduction to MultiAgent Systems” Appendix A